{"id":"dc1a6b7f-f57d-480a-9960-f8db5d408127","arxiv_id":"2501.14185","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A tensor-based binary encoding maps graph vertices and edges to Pauli-Z strings for variational quantum classifiers, and simulations on three datasets show accuracy gains over a PCA-based baseline, with several unaddressed theoretical gaps.","lead":"The authors propose a new way to encode graph data into quantum circuits for classification, using binary vertex indices to build Pauli-Z operator strings. They report accuracy gains over a PCA-based baseline on three small benchmark datasets, but the paper contains serious theoretical and implementation gaps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The encoding in Eqs. (6)-(8) assigns Pauli strings from arbitrary vertex indices, so relabeling vertices changes the Hamiltonian and can change VQC predictions; graph-information preservation is therefore unestablished on the benchmark datasets.","rationale":"The reader's weakest assumption points to the same place, and I agree that it is the decisive issue. The central claim is not merely that the circuit is implementable or that the appendix bound is stated correctly; it is that EG-VQC captures graph structure well enough to classify graphs. A representation that depends on an arbitrary integer labeling of vertices does not satisfy the most basic invariance property expected of a graph classifier, and the paper never states a canonical labeling scheme. The appendix proof issue is real but secondary: the bound is conditional on Delta_J + Delta_h <= 1 and is not itself the basis of classification. The lack of a concrete circuit implementation is also serious for reproducibility, but it is a completeness problem rather than a conceptual invalidation of the method; if the permutation issue is confirmed, no amount of implementation detail can save the claimed graph-preservation property. My recommendation is to leave the reader's REJECT verdict unchanged.","tokens_in":9579,"tokens_out":5034,"duration_ms":50631,"concrete_test":"Take a trained EG-VQC model (or re-train using the paper's setup) on one dataset. For a fixed validation graph, generate 50 random vertex relabelings pi, reorder its adjacency matrix as A_pi = P A P^T, encode each relabeled graph via Eqs. (6)-(8), and record the predicted class probabilities with the same trained parameters and no retraining. If predictions vary across relabelings, the encoding is not permutation-invariant; report the variance and the change in test accuracy when a held-out set is randomly relabeled. A second, analytic check is to verify whether H(pi(G)) = U_pi H(G) U_pi^dagger for a fixed family of unitaries; for a relabeling that is not affine over GF(2)^N this equality fails, settling the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the headline claim that EG-VQC 'preserves the integrity of graph data,' the encoding must assign equivalent (or at least invariant) representations to isomorphic graphs, since MUTAG, PROTEIN, and ENZYME labels are invariant under vertex relabeling. The paper does not provide this. Equation (6) encodes vertex i using the binary expansion of the integer i, and Eq. (8) builds H(G) from these binary-indexed strings. If vertices are relabeled by a permutation pi, the set of Pauli strings changes; H(pi(G)) is generally not unitarily equivalent to H(G) via a fixed N-qubit transformation, unless pi is a linear map on GF(2)^N. No canonical vertex ordering is imposed in Algorithm 1, and Section III-C2 explicitly states that the method 'eliminates the need for strict vertex alignment,' which is the opposite of what the encoding requires. As a result, the same graph can be encoded as different Hamiltonians and the trained VQC can return different class probabilities depending on the arbitrary ordering supplied by the dataset. The observed accuracy gains in Table I may therefore reflect the particular ordering used in preprocessing rather than graph structure. This is a representation-level failure of the central claim, independent of the implementation details and the appendix eigenvalue-bound issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a graph encoding scheme for variational quantum classifiers, called EG-VQC. Each vertex is mapped to a binary-indexed tensor product of Pauli-Z operators (Eq. 6) and each edge to a weighted product of such operators (Eq. 7), so that a graph becomes a Hamiltonian H(G) of Eq. (8). The authors argue that this encoding preserves graph information, avoids PCA information loss, and requires fewer qubits than PCA-VQC. They present numerical results on MUTAG, PROTEIN, and ENZYME (Table I) claiming higher classification accuracy than PCA-VQC, together with a complexity analysis and an appendix proof intended to show that H(G) can be normalized to have eigenvalues in [-1,1]. The central claims are that the encoding preserves the integrity of graph data and that the resulting classifier outperforms PCA-VQC.","tokens_in":9887,"tokens_out":3473,"duration_ms":34878,"significance":"If the claims were correct, a permutation-robust, qubit-efficient graph encoding for VQCs would be a useful contribution to quantum machine learning. The paper also makes a potentially valuable point: full graph information can be encoded without PCA dimensionality reduction, and the O(N_G^2) encoding complexity is an improvement over the O(N_G^3) SVD step. The numerical comparison with PCA-VQC is a plausible and falsifiable target. However, the manuscript as written does not establish the central claims: the encoding is not permutation invariant, the eigenvalue proof is mathematically incorrect, and the algorithm contains an undefined measurement formula. These issues are load-bearing for the paper's main assertions, so the current version is not publishable.","major_comments":[{"comment":"The encoding is not invariant under vertex relabeling, which is a fundamental problem for graph classification. Equation (6) assigns the Pauli string based on the binary expansion of an integer label i, so a graph and any of its isomorphic copies with a different vertex ordering produce, in general, different Hamiltonians and different VQC predictions. Section III-C2 explicitly states that the method 'eliminates the need for strict vertex alignment,' but the opposite is true: without a canonical vertex ordering, the same unlabeled graph can yield different classification outcomes depending on the arbitrary order in which its vertices appear in the dataset. Since MUTAG, PROTEIN, and ENZYME are standard permutation-invariant graph benchmarks, this representation-level flaw undermines the claim that EG-VQC 'preserves the integrity of graph data.'","section":"III-A, Eq. (6)-(8); III-C2"},{"comment":"The eigenvalue bound is incorrect. The appendix states that because each Pauli term has eigenvalues ±1 and coefficients J_ij and h_i are multiplied with these terms, 'the eigenvalues of H(G) are bounded by the maximum of the absolute values of J_ij and h_i,' and then concludes λ(H(G)) ∈ [-(ΔJ+Δh), ΔJ+Δh]. This is false for a sum of operators. Since all terms in Eq. (8) are diagonal in the computational basis, the eigenvalue for a basis state |b⟩ is Σ_{ij∈E} ±J_ij + Σ_i ±h_i, whose absolute value can be as large as Σ|J_ij| + Σ|h_i|, not merely ΔJ+Δh. The normalization condition ΔJ+Δh≤1 is therefore not sufficient to guarantee eigenvalues in [-1,1]. The boundary condition is imposed as a construction choice, not derived, so the proof in the appendix does not support the stated claim.","section":"Appendix A.4, combined contribution"},{"comment":"The measurement formula 'pi = ⟨0|H(G)U(θ)|0⟩' is not a valid probability. H(G) is a Hermitian operator, not a positive operator-valued measure element, and the expression is not guaranteed to be real-valued nonnegative or bounded in [0,1]. Moreover, the order of H(G) and U(θ) is unclear: if the Hamiltonian is used to encode the graph, the state should be evolved or measured in a way that the encoding appears before the variational circuit. As written, the cross-entropy loss in Section III-C cannot be computed from this formula. This is a load-bearing implementation detail because the entire numerical section depends on it.","section":"Algorithm 1, line 15"},{"comment":"The central empirical claim rests entirely on Table I, but the manuscript provides no details about the experimental protocol beyond '90:10 split with stratified sampling.' It does not state the number of independent runs, the hyperparameters of the PCA-VQC baseline, the circuit depth used for each dataset, the optimizer settings beyond learning rate and epochs, or how the reported means and standard deviations were computed. Without these details, the accuracy differences (3-8 percentage points) cannot be assessed for statistical significance. The absence of code or a detailed methodology makes the results non-reproducible.","section":"IV-C, Table I"}],"minor_comments":[{"comment":"There is a typo in the section heading: 'PAC-VQC' should be 'PCA-VQC.'","section":"II.D"},{"comment":"Reference [28] is cited as the source of the MUTAG dataset, but the listed paper (Debnath, Soni, and Reddy, 'A survey on graph neural networks') is not the MUTAG dataset reference. The MUTAG dataset should be cited to its original source, e.g., Debnath et al., J. Med. Chem. 1991.","section":"References"},{"comment":"The initialization is inconsistent: line 10 says to initialize qubits with Hadamard gates, but the measurement formula in line 15 uses the state |0⟩. The paper should clarify which initial state is used and how the Hamiltonian encoding in Eq. (8) is converted into a circuit that acts on that state.","section":"Algorithm 1, line 10"},{"comment":"The caption of Fig. 3 states that the training process using MUTAG and PROTEIN is shown, but the figure is not referenced in the main text and no loss curves are described quantitatively. Please clarify what is plotted and how the loss trends support the claim of reduced overfitting.","section":"IV-C2, Fig. 3"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an early-stage draft with several internal inconsistencies and an incorrect mathematical proof. The permutation-invariance issue is fundamental to the proposed encoding and cannot be fixed by minor edits; it would require a redesigned representation or an explicit canonicalization procedure. The empirical section is also too thin to support the headline accuracy claims. I recommend rejection, though the general idea of encoding graphs as weighted Pauli strings is worth exploring in a more rigorous follow-up."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my read. The core idea is a new way to encode graphs into a Hamiltonian-like operator using binary-indexed Pauli strings, giving O(log N) qubits. That part is genuinely new and the complexity story is plausible. The empirical tables show a modest but consistent improvement over the PCA-VQC baseline, and the paper is honest that classical GNNs beat both. If I stopped there, this would be a decent niche result.\n\nBut the load-bearing claim—that the encoding preserves graph integrity—does not survive contact with the actual encoding. Eq. (6) assigns each vertex a Pauli string from its integer index. Relabeling vertices changes the Hamiltonian in a way that is not equivalent under any fixed unitary if the permutation is arbitrary. Since the benchmark datasets are order-invariant, the reported accuracies may simply reflect a particular vertex ordering. Section III-C2 claims the method eliminates the need for strict vertex alignment, which is exactly backwards. This is a real representation-level flaw and it undercuts the central message.\n\nThere are also two technical problems. Algorithm 1's measurement formula pi = <0|H(G)U(theta)|0> is not a probability and no conversion is specified. The appendix's eigenvalue proof claims the eigenvalues are bounded by the max coefficient, but for a sum of commuting Pauli operators the bound involves the sum of absolute coefficients; the stated condition does not ensure the range [−1,1]. Both are fixable in principle.\n\nWhat the paper does well: the encoding is new, the complexity comparison is sensible, and the experiments are reproducible in principle (PennyLane) though no code is provided. The flaws are not fabricated—the permutation issue is decisive.\n\nMy recommendation: this deserves serious peer review, not because it is near acceptance but because the core construction is novel and a careful referee can help reshape it. It currently needs major revision: specify a canonical ordering (or honestly restrict the setting), provide circuit details and code, and correct the bound. I would not cite this in its current form. It might be a good reading group discussion piece, though more to illustrate what can go wrong with quantum graph encodings than as a positive result.","headline":"Novel binary Pauli-string graph encoding undercut by a permutation-invariance flaw, an under-specified circuit, and an incorrect eigenvalue bound.","tokens_in":10358,"tokens_out":3825,"would_cite":false,"duration_ms":37660,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.Lx"],"model":"deepseek-v4-flash","headline":"The paper claims that encoding each graph vertex as a binary-indexed Pauli-Z tensor product and each edge as a weighted product preserves full graph structure, allowing a variational quantum classifier to beat the PCA-based baseline on…","keywords":["variational quantum classifier","graph classification","Pauli-Z encoding","tensor product encoding","Ising model","quantum machine learning","NISQ","graph encoding"],"falsifier":"Train the same EG-VQC circuit on the MUTAG, PROTEIN, or ENZYME datasets twice: once with the original vertex order and once with a random permutation of vertex indices applied consistently across all graphs (so adjacency is unchanged). If test accuracy drops materially under relabeling, the encoding is not capturing permutation-invariant graph structure; if accuracy is unchanged, the claim that the binary-indexed encoding itself carries the structural information is supported.","tokens_in":9383,"feed_emoji":"⚛️","tokens_out":13915,"duration_ms":93173,"temperature":0.7,"pith_summary":"The paper proposes a graph encoding for variational quantum classifiers (VQCs) in which each vertex is mapped to a tensor product of Pauli-$Z$ operators determined by the vertex's binary index, and each edge is mapped to a weighted product of the two vertex strings, forming an Ising-model-like Hamiltonian. The authors claim this encoding preserves the full graph structure, unlike PCA-based preprocessing which discards information, and it requires only $O(\\log N_G)$ qubits. They report that their EG-VQC achieves higher accuracy than PCA-VQC on the MUTAG, PROTEIN, and ENZYME datasets (e.g., 80.3% vs 75.1% on MUTAG), with lower time complexity. A sympathetic reader would care because it suggests a practical, NISQ-friendly way to feed graph data into variational quantum circuits without dimensionality reduction.","feed_headline":"Graphs as Pauli strings beat PCA on quantum classifiers","feed_subtitle":"This encoding of vertices and edges as weighted Pauli-Z tensors preserves graph structure and beats PCA accuracy.","key_machinery":"The enabling object is the binary-indexed Pauli-$Z$ tensor product: for a vertex labeled $i$ with binary expansion $\\sum_{k=0}^{N-1} r_k 2^k$, the vertex is represented by $\\bigotimes_{k=0}^{N-1} Z^{r_k}$. Edges are represented by weighted products $Z_i^{\\otimes} Z_j^{\\otimes}$, producing an Ising-model-style Hamiltonian whose expectation values are measured by the VQC. This tensor encoding is what carries the graph structure into the quantum circuit, and the eigenvalue bound lemma in the appendix is what keeps the resulting observable within a range compatible with the cross-entropy loss.","core_discovery":"The central claim is that a graph can be encoded into a quantum Hamiltonian of the form $$H(G) = \\sum_{\\{i,j\\}\\in E} J_{ij} $Z_i^{{\\otimes}}$ $Z_j^{{\\otimes}}$ + \\sum_{i\\in V} h_i $Z_i^{{\\otimes}}$,$$ where $Z_i^{\\otimes}$ is the tensor product of Pauli-$Z$ operators indexed by the binary representation of vertex $i$, and $J_{ij}, h_i$ are normalized weights. The paper argues that this encoding retains the complete interaction structure of the graph, that its eigenvalues can be bounded to $[-1,1]$ (with proof in the appendix under the condition $\\Delta J + \\Delta h \\le 1$), and that feeding this Hamiltonian into a VQC with variational $RY$ and $CNOT$ layers yields better classification accuracy than the PCA-VQC baseline on the three benchmark datasets.","pith_inferences":["The encoding as written is tied to the arbitrary integer labels of vertices; relabeling vertices produces a different Hamiltonian, so the method as presented is not permutation-invariant, which is the usual setting for graph classification benchmarks. A permutation-invariant variant would need a canonical vertex ordering or a symmetrization over labelings.","Because the full graph Hamiltonian is used as the observable, the method might also be read as a quantum kernel: two graphs can be compared by their Hamiltonian expectation values, suggesting a possible route to quantum graph kernels that avoid classical feature extraction.","A direct testable extension is to run the same pipeline on larger graph benchmarks (e.g., NCI1, DD) to see whether the accuracy gap over PCA-VQC persists as graph size and sparsity vary.","The eigenvalue-bound proof in the appendix only covers the special case $\\Delta J + \\Delta h \\le 1$; for general graphs the Hamiltonian may leave the $[-1,1]$ interval, so future work would need a normalization strategy that holds unconditionally."],"forward_implications":["Graph structure can be fed into a VQC without PCA-style dimensionality reduction, eliminating a known source of information loss.","The $O(N_G^2)$ encoding time beats the $O(N_G^3)$ SVD step of PCA-VQC, so larger graphs become feasible on limited hardware.","The method needs only $\\lceil \\log_2 N_G \\rceil$ qubits, so graphs with up to $2^N$ vertices can be processed on an $N$-qubit device.","On the tested benchmarks, EG-VQC yields higher validation accuracy and lower overfitting (flatter validation loss) than PCA-VQC.","The encoding is compatible with cross-entropy loss and standard variational training (Adam over 100 epochs)."],"supporting_citations":[{"why":"Supplies the theoretical basis for variational hybrid quantum-classical algorithms that VQCs build on.","marker":"[1]"},{"why":"Defines the circuit-centric quantum classifier architecture that the proposed encoding plugs into.","marker":"[9]"},{"why":"Provides the PCA-VQC baseline that the experiments compare against; the proposed method is benchmarked against this approach.","marker":"[20]"},{"why":"Gives the exact Ising-model simulation formalism that motivates the tensor and Pauli-Z encoding of graph interactions.","marker":"[21]"},{"why":"Provides the quantum simulation library used for all numerical experiments in the paper.","marker":"[25]"},{"why":"Supplies the MUTAG dataset of mutagenic compounds used in the accuracy comparison.","marker":"[28]"},{"why":"Supplies the PROTEIN dataset used in the accuracy comparison.","marker":"[29]"},{"why":"Supplies the ENZYME dataset used in the accuracy comparison.","marker":"[30]"}],"fun_headline_variants":["Quantum graph encoding beats PCA in classification","Pauli tensor graph encoding surpasses PCA-VQC","NISQ-friendly graph encoding outperforms PCA baseline","Graph-as-Pauli-tensor encoding beats PCA for NISQ","Preserving graph structure, tensor encoding beats PCA in VQC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The encoding takes the arbitrary numbering of vertices as meaningful: permuting the vertex labels changes the Hamiltonian and therefore the learned classifier, so the method as presented is not invariant under graph isomorphism, the usual symmetry of graph classification.","fun_headline_variants_meta":{"raw":{"variants":["Quantum graph encoding beats PCA in classification","Pauli tensor graph encoding surpasses PCA-VQC","NISQ-friendly graph encoding outperforms PCA baseline","Graph-as-Pauli-tensor encoding beats PCA for NISQ","Preserving graph structure, tensor encoding beats PCA in VQC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000713,"raw_usage":{"total_tokens":3226,"prompt_tokens":984,"completion_tokens":2242,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":2163}},"tokens_in":600,"tokens_out":2242,"duration_ms":13742,"temperature":1.0,"reasoning_tokens":2163,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:15:15.705354+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same EG-VQC circuit on the MUTAG, PROTEIN, or ENZYME datasets twice: once with the original vertex order and once with a random permutation of vertex indices applied consistently across all graphs (so adjacency is unchanged). If test accuracy drops materially under relabeling, the encoding is not capturing permutation-invariant graph structure; if accuracy is unchanged, the claim that the binary-indexed encoding itself carries the structural information is supported.","supporting_citations":[{"cited_title":"Hybrid quantum-classical classifier based on tensor network and variational quantum circuit,","cited_arxiv_id":null,"evidence_quote":"Provides the PCA-VQC baseline that the experiments compare against; the proposed method is benchmarked against this approach."},{"cited_title":"A survey on graph neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the MUTAG dataset of mutagenic compounds used in the accuracy comparison."},{"cited_title":"Protein classification with graph kernels,","cited_arxiv_id":null,"evidence_quote":"Supplies the PROTEIN dataset used in the accuracy comparison."},{"cited_title":"Enzyme classification in terms of graph-based representations,","cited_arxiv_id":null,"evidence_quote":"Supplies the ENZYME dataset used in the accuracy comparison."}],"review_version":1}